Generative Inpainting of Partially destroyed Frescos

Abstract. Fresco painting, a prevalent mural technique, is highly susceptible to damage due to the hygroscopic nature of lime plaster, often resulting in partially destroyed artworks. Restoring these masterpieces poses significant challenges: merging remaining fragments with new additions requires skilled restoration, while reconstructing original compositions demands expert historical insight. The absence of textual or graphical records further complicates this task. Recent advancements in neural inpainting offer potential solutions but lack the precision required by art historians. We introduce the Stable Restorer model, enhancing neural inpainting with fine-grained control over generated content. Building on the Flux model with LoRA adaptation, our approach employs multiple text prompts linked to attention masks for precise local editing. Our Mural Paintings dataset, comprising 15k samples of diverse frescos, facilitates rigorous evaluation. Results indicate that our model not only competes with but surpasses state-of-the-art methods, offering art historians enhanced control over reconstructed fresco regions.